An intelligent detection method for precise analysis of shield tunnel lining joints based on deep learning networks and image morphology algorithms

Leakage between adjacent lining blocks can cause severe secondary accidents, necessitating prompt joint inspection and repair. This study introduces an innovative framework to evaluate leakage of tunnel lining joints quantitatively. First, an enhanced two-stage detection model is proposed and utilised to predict joint bounding boxes. A novel backbone integrating Swin Transformer and sandglass blocks is used to capture both holistic and regional features. After the leakage joints are detected and located, leakage in the images is segmented based on the threshold algorithm, where a process containing image morphology algorithms is designed to enhance the quality and reasonability of the leakage masks. According to the characteristics of shield tunnels, the quantitative assessment method is proposed to obtain the areas of leakage regions. The mAP_50-95 of the proposed detector improves by 5.7% to 8.9% compared to previous methods, while the FPS reaches 15.68. The proposed method performs well on images larger than 400 × 400 pixels, and the AP₅₀ can exceed 80%. For the segmentation method, most IoUs between the manual and segmented masks are larger than 0.7. Thresholds specified in the Chinese standards are used to evaluate the leakage condition of an on-site tunnel, and maintenance guidance is given. Highlights Enhanced Faster R-CNN is proposed to improve the performance of detecting lining joints. Threshold and image morphology algorithms are combined to obtain reasonable leakage segmentation masks. A quantified algorithm is designed to calculate the pixel sizes for obtaining the actual areas of the leakage masks. The performance of the proposed method is evaluated based on images from a tunnel section containing 664 rings.